MIT researchers report an open-source machine-learning layer that learns recurring errors in existing weather models and corrects their probabilistic forecasts two to six weeks ahead. In an ECMWF competition, a system using the wrapper doubled the base model’s predictive power, beat climatological averages in that difficult range and improved extreme-weather forecasts; a conventional mean-seeking bias correction had made extremes worse. The useful idea is modular: improve an established dynamical model’s residuals instead of pretending an AI forecast should replace the physics stack. It does not make a six-week local forecast deterministic. MIT News, Oct. 14